At its core, ACE-Step 1.5 has a novel hybrid architecture where the Language Model functions as an omni-capable planner, transforming simple user queries into comprehensive song blueprints. It synthesizes metadata, lyrics, and captions via Chain-of-Thought to guide the Diffusion Transformer, achieving alignment through intrinsic reinforcement learning. This eliminates biases inherent in external reward models or human preferences, enabling precise stylistic control and versatile editing capabilities.
ACE-Step 1.5 unifies precise stylistic control with versatile editing capabilities, such as cover generation, repainting, and vocal-to-BGM conversion, while maintaining strict adherence to prompts across 50+ languages. The model has been compared to other commercial and open-source music generation models, demonstrating its efficiency and quality. However, it also has some limitations, including output inconsistency, style-specific weaknesses, and continuity artifacts, which are being addressed for future improvements.
